27B decision model that returns calibrated probabilities for routing, classification, and scoring instead of generating text.
Output Speed *
Intelligence Index *
Context Window *
Input price
Output price
Performance
Top Jev Decision Index Score at Release, With a Thin Margin on the Artificial Analysis Replica
Technical Specifications
What the model supports
Comparison
mammoth-1 or elephant-1: Which Decision Model Belongs in Your Routing Layer?
Base Labs released elephant-1 and mammoth-1 together on October 21, 2026. They share the same architecture, training data, and System One request format. The difference is scale: 27B (Qwen3.5-27B) versus 9B (Qwen3.5-9B). On Jev 0.4, mammoth-1 scores 71.24 against roughly 62.6 for elephant-1. On Base Labs' Artificial Analysis replica the two are close, at 65.9 and 65.2. The practical decision is broader knowledge and higher Jev accuracy versus faster, lower-cost decisions.
Choose mammoth-1 when tasks depend on general knowledge, when classification is ambiguous or knowledge-heavy, or when you want the strongest Jev-measured accuracy. Choose elephant-1 when decisions are simple, latency and cost per decision dominate, and your own tests show little accuracy loss. Baseten recommends it for quick decisions. Because the replica gap is small, run both on your hardest cases before defaulting to the larger model.
Limitations & Trade-offs
Best-Fit Workloads
Where this model earns its place
Support ticket triage and routing
One request can choose the owning team, test refund intent, and rate urgency on a scale you define. Baseten's own example routes a ticket automatically when the selected option's probability is at least 0.9 and sends it to review otherwise. mammoth-1 suits queues where tickets need broader knowledge to classify. It is a poor fit if agents also need a drafted reply, since it writes no text.
Confidence-gated escalation
mammoth-1 returns a probability for every option, so code can accept high-confidence answers and escalate uncertain ones to a human. Training included examples that teach the model when to answer "I don't know." This helps review queues and approval flows. Calibration should still be checked on your own data, because the published results are vendor-reported.
Guardrail and safety checks
Baseten positions the mammoth-1 and elephant-1 pair for tasks ranging from ticket routing to checking requests for safety, and training data included safety and abstention examples. Yes/no probabilities let you set pass, review, or block thresholds ahead of a generative model. For very high-volume, latency-sensitive guardrails, compare it with the smaller elephant-1 or joey-1.
Rubric-based content scoring
Ordinal score questions rate content against criteria you write, and a score can fall between rubric levels. State can include images, which suits moderation or quality review of mixed text and image content. Use it for consistent numeric grading at scale. It will not explain a score, so keep a generative judge for cases needing written justification.